A2A (Agent-to-Agent Protocol) is Google’s open standard for AI agents to communicate with each other. While MCP connects agents to tools (vertical), A2A connects agents to other agents (horizontal).
Google launched A2A in April 2025 with 50+ enterprise partners including Salesforce, SAP, and ServiceNow. It’s now under the Linux Foundation alongside MCP.
How it works
A2A defines four concepts:
Agent Cards — JSON descriptions of what an agent can do (like a business card for AI)
Tasks — Units of work that agents delegate to each other
Messages — Structured communication between agents
Artifacts — Files and data shared between agents
Agent A discovers Agent B's capabilities (Agent Card)
→ Agent A creates a Task and sends it to Agent B
→ Agent B processes, sends status updates
→ Agent B returns results as ArtifactsA2A vs MCP
They’re complementary, not competing:
| A2A | MCP | |
|---|---|---|
| Connects | Agent ↔ Agent | Agent ↔ Tools |
| Example | Support agent → Billing agent | Agent → Database |
| When | Multi-agent workflows | Tool integration |
Most production systems use both. See our full comparison.
Key concepts
Agent Cards — JSON descriptions of what an agent can do. Like a business card for AI — other agents discover capabilities by reading the card.
Tasks — Units of work delegated between agents. An agent creates a task, sends it to another agent, and receives results.
Messages — Structured communication during task execution. Status updates, clarifying questions, intermediate results.
Artifacts — Files and data shared between agents as task outputs.
Example flow
1. Support agent receives customer complaint about billing
2. Support agent reads Billing Agent's Agent Card
3. Support agent creates a Task: "Process refund for order #12345"
4. Billing agent processes the refund, sends status updates
5. Billing agent returns confirmation as an Artifact
6. Support agent tells the customer it's doneWho uses A2A?
Google launched A2A with 50+ enterprise partners including Salesforce, SAP, ServiceNow, and Atlassian. It’s now under the Linux Foundation alongside MCP, co-governed by OpenAI, Google, Microsoft, and Anthropic.
For developers
If you’re building single-agent tools (Claude Code, Aider), focus on MCP first. A2A matters when you need multi-agent coordination like Kimi’s Agent Swarm.
FAQ
When should I use A2A instead of just calling another service’s API?
Use A2A when you need AI agents to dynamically discover each other’s capabilities, delegate complex tasks, and exchange structured results — especially when the agents are built by different teams or organizations. For simple service-to-service calls with fixed contracts, a regular API is simpler and sufficient.
Can A2A work with agents built on different LLM providers?
Yes — A2A is provider-agnostic. An agent powered by Claude can delegate tasks to an agent powered by GPT or Gemini. The protocol standardizes the communication format (Agent Cards, Tasks, Messages) regardless of the underlying model or framework.
Do I need both MCP and A2A?
Not necessarily — it depends on your architecture. If you’re building a single agent that needs tool access, MCP alone is enough. A2A becomes valuable when you have multiple specialized agents that need to coordinate. Most complex production systems use both: MCP for tool integration and A2A for agent-to-agent delegation.
Related: MCP vs A2A vs ACP · What is MCP? · MCP Complete Guide